EdgeTypE/Gokturk_OCR_Test

Work in progress.

Gokturk ResNet OCR Model

This is a ResNet-based OCR model specifically trained to recognize Old Turkic (Gokturk) script characters. It is optimized for inference using ONNX Runtime, making it highly portable and efficient.

Model Description

• Repository: OldTurkicOCR
• Task: Optical Character Recognition (OCR) / Image Classification
• Classes: 75 characters (Unicode range U+10C00 – U+10C4A)
• Input Shape: [(Batch, 64, 64, 1)](file:///a:/Users/Edige/GitHub/ocr-gokturk/src/ocr.rs#36-60) (Grayscale)
• Format: ONNX

How to use

This model is primarily used within the OldTurkicOCR Rust project.

In Rust (with ort crate)

use ort::session::Session;

let session = Session::builder()?
    .commit_from_file("gokturk_resnet_v1.onnx")?;

In Python (with onnxruntime)

import onnxruntime as ort
import numpy as np

session = ort.InferenceSession("gokturk_resnet_v1.onnx")
# Expects a 64x64 grayscale image normalized to [0, 1]
# input_data shape: (1, 64, 64, 1)
result = session.run(None, {"input_1": input_data})

Dataset and Training

The model was trained on a curated dataset of Gokturk script characters, covering various styles and weights of the Orkhon and Yenisei variants.

Files

• [gokturk_resnet_v1.onnx](file:///a:/Users/Edige/GitHub/ocr-gokturk/gokturk_resnet_v1.onnx): The main inference model.
• [model_labels.json](file:///a:/Users/Edige/GitHub/ocr-gokturk/model_labels.json): Mapping of model output indices to character labels.